Research And Implementation Of Automatic Evaluation System For Embryo Quality In Pronuclear Stage

Assisted Reproductive Technology(ART)refers to the use of medical assistive technology to infer the infertile couple’s technology,including Artificial Insemination(AI)and In Vitro Fertilization and Embryo Transfer(IVF-ET).And its derivative technology,in which in vitro fertilization-embryo transfer is the core of the entire assisted reproductive technology.After fertilization,the embryo develops in vitro through the pronuclear stage,cleavage stage,and blastocyst stage.It is generally used to transplant the embryo into the mother in the blastocyst stage.The higher the quality of embryo development before transplantation,the higher the success rate of pregnancy,so screening high-quality embryos has become the core task of embryo culture in vitro.Traditionally,in the process of in vitro embryo culture,the embryonic biocharacter is recorded by experienced embryologists at various stages of embryonic development,and then the developmental quality of the embryo is judged according to standard embryo evaluation criteria.Due to the different experience of each embryologist,the final recorded embryo biometrics will inevitably be biased,resulting in the final assessment results subjective.The rise of deep learning has made it possible for computer systems to replace embryologists in accurately capturing biological features during embryonic development.The research content of this thesis is based on the sub-project under the research and implementation of the scientific research project “Automated Embryo Development Monitoring System” jointly developed by a top three hospital in Chengdu.The author is the main researcher of this sub-project.Based on the deep learning model,this paper detects and identifies the biological characteristics of pronuclear embryos,including prokaryotic and transparent bands.Then combined with the general evaluation criteria,construct a deep learning model,learn the evaluation level of biometrics,and finally make a comprehensive evaluation of the embryo quality in the pronuclear stage.There is very little literature on the use of deep learning for in vitro development of embryos,which is basically absent in China.Researchers abroad have used deep learning techniques to allow computers to automatically detect and identify various physiological features during embryonic development,and to analyze their morphological characteristics by computer,but these studies are mostly directed at blastocyst stage embryos or cleavage stage embryos.There is very little research literature on pronuclear embryos.The main research contents of this thesis are as follows:(1)Based on the YOLOv2 model,the embryo target was located in the pronuclear embryo;then the FCN semantic segmentation model was used to classify the embryonic biometrics in the pronuclear stage to achieve the purpose of extracting biometrics separately;then the VGG-16 model was trained to perform the prokaryotic level.Classification,predicting the number of pronuclear,and completing the integrity judgment of the transparent zone by designing the algorithm;finally,evaluating the quality of the pronuclear embryo.(2)Due to the small amount of image data in the pronuclear embryo,in order to meet the requirements of a large amount of data for deep learning model training,and to improve the generalization ability of the model,it is necessary to enhance the data of the existing data.This paper combines the commonly used image transformation.The method designs a data enhancement algorithm.(3)This paper has designed the automatic evaluation system of embryo quality in the pronuclear stage,and divided the whole system into four functional modules,data enhancement module,pronuclear embryo positioning module,pronuclear embryo biometric extraction module,biometrics and The comprehensive evaluation module of embryo quality in the pronuclear stage elaborated the models and algorithms involved in each functional module,and finally completed the implementation of the whole system.The main innovations of this paper are as follows:(1)Applying YOLOv2 target detection technology to embryo localization,improve the efficiency and accuracy of embryo localization by adjusting the YOLOv2 model.(2)Improve the sampling process of FCN semantic segmentation model,add learning weight network in the upsampling process,make the characteristics of the upper layer more affect the network characteristics of the lower layer,and improve the accuracy and robustness of biometric segmentation.(3)By designing the VGG-16 model to classify the pronuclear level and the number of pronuclear,the classification of the pronuclear level is more accurate.(4)The data enhancement algorithm is designed to make different image transformation methods randomly combined.The images are more diversified after data enhancement,which is conducive to the generalization ability of deep learning models.Experimental tests show that the deep learning model used in this paper can accurately perform embryo localization and biometric extraction.Through comparison experiments,the effectiveness of the improved FCN semantic segmentation model is verified,and the system can be tested to coordinate various functional modules,working normally.